XPENG–Peking College Collaborative Analysis Accepted By AAAI 2026: Introducing A Novel Visible Token Pruning Framework For Autonomous Driving

Assist CleanTechnica’s work by means of a Substack subscription or on Stripe.

• XPENG-PKU Analysis Breakthrough: XPENG, in collaboration with Peking College, has developed FastDriveVLA—a novel visible token pruning framework that allows autonomous driving AI to “drive like a human” by focusing solely on important info, attaining a 7.5x discount in computational load.
• High-Tier AI Recognition: The analysis has been accepted by AAAI 2026, one of many world’s premier AI conferences, which had a extremely selective acceptance fee of simply 17.6% this 12 months.
• Accelerating L4 Autonomy: This achievement underscores XPENG’s full-stack capabilities in AI-driven mobility and advances the trade towards environment friendly, scalable deployment of next-generation autonomous driving techniques.

GUANGZHOU, China – XPENG, in collaboration with Peking College, has had its paper “FastDriveVLA: Environment friendly Finish-to-Finish Driving through Plug-and-Play Reconstruction-based Token Pruning” accepted by AAAI 2026, one of many world’s prime conferences in synthetic intelligence. AAAI 2026 acquired 23,680 submissions, with solely 4,167 papers accepted, an acceptance fee of simply 17.6%.

The paper introduces FastDriveVLA, an environment friendly visible token pruning framework particularly designed for end-to-end autonomous driving Imaginative and prescient-Language-Motion (VLA) fashions. This work presents a brand new method to visible token pruning by enabling AI to “drive like a human”, focusing solely on important visible info whereas filtering out irrelevant knowledge.

As AI giant fashions evolve quickly, VLA fashions are being extensively adopted in end-to-end autonomous driving techniques attributable to their sturdy capabilities in advanced scene understanding and motion reasoning. These fashions encode photos into giant numbers of visible tokens, which function the muse for the mannequin to “see” the world and make driving choices. Nevertheless, processing giant numbers of tokens will increase computational load onboard the car, impacting inference pace and real-time efficiency.

Whereas visible token pruning has been acknowledged as a viable methodology to speed up VLA inference, present approaches, whether or not based mostly on text-visual consideration or token similarity, have proven limitations in driving situations. To handle this, XPENG and PKU developed FastDriveVLA, a novel reconstruction-based token pruning framework impressed by how human drivers deal with related foreground info (e.g., lanes, autos, pedestrians) whereas ignoring non-critical background areas.

The strategy introduces an adversarial foreground-background reconstruction technique that enhances the mannequin’s skill to determine and retain beneficial tokens. On the nuScenes autonomous driving benchmark, FastDriveVLA achieved state-of-the-art efficiency throughout varied pruning ratios. When the variety of visible tokens was decreased from 3,249 to 812, the framework achieved an almost 7.5x discount in computational load whereas sustaining excessive planning accuracy.

That is the second time this 12 months that XPENG has been acknowledged at a top-tier international AI convention. In June, XPENG was the one Chinese language automaker invited to talk at CVPR WAD, the place it shared advances in autonomous driving basis fashions. At its Tech Day in November, XPENG unveiled its VLA 2.0 structure, which removes the “language translation” step and allows direct Visible-to-Motion era, a breakthrough that redefines the standard V-L-A pipeline.

These accomplishments replicate XPENG’s full-stack in-house capabilities, from mannequin structure design and coaching to distillation and car deployment. Trying forward, XPENG stays dedicated to attaining L4 stage autonomous driving. We are going to proceed to put money into AI giant mannequin expertise to speed up the combination of bodily AI techniques into autos, with the objective of delivering protected, environment friendly, and comfy clever driving experiences to customers all over the world.

About XPENG

XPENG is dedicated to main the transformation of future mobility by means of technological exploration, positioning itself as “Explorer of Future Mobility”. Headquartered in Guangzhou, China, the corporate operates R&D facilities in Beijing, Shanghai, Shenzhen, Zhaoqing, and Yangzhou, and has established clever manufacturing bases in Zhaoqing and Guangzhou.
XPENG pursues a world technique for analysis, growth, and gross sales, with an R&D heart in the US and subsidiaries throughout a number of European nations. The corporate adheres to full-stack in-house growth of clever driver-assistance software program and the event of core {hardware}, delivering an distinctive clever driving and using expertise for customers.
On August 27, 2020, XPENG formally listed on the New York Inventory Change (NYSE: XPEV), elevating funds in an IPO that set a document on the time for the worldwide new vitality car trade. On July 7, 2021, the corporate listed on the Hong Kong Inventory Change (HKEX: 9868), turning into the primary Chinese language new-energy automaker to attain twin major listings in each Hong Kong and New York.

For extra info, please go to https://www.xpeng.com/.